DocsHaven

Your AI agent keeps forgetting what it learned last session. DocsHaven fixes that.

Add your repos, and your agent always has them at hand. DocsHaven is a local knowledge base that lets you index any GitHub repository, search it instantly, and keep your agent informed across sessions. Works with any MCP-compatible agent — Claude, Cursor, Gemini, Codex.

Why DocsHaven?

• 🧠 Add repos once, search forever — your agent always has the knowledge it needs
• ⚡ Instant search — SQLite FTS5 finds relevant docs in <100ms
• 🔌 Works with any AI agent — Claude, Cursor, Gemini, Codex via MCP
• 🏠 100% local — no cloud, no API keys, no data leaves your machine
• 📦 Minimal dependencies — only mcp[cli], no Docker or external services

Features

• SQLite FTS5 search — BM25 ranking with LIKE fallback, score explanation
• URI routing — organize knowledge by domain: core://, ref://, guide://
• Git sync — compressed chunks for multi-machine sync (no merge conflicts)
• Conflict detection — flag contradictions when adding documents
• MCP server — 20 tools for any MCP-compatible agent
• Document chunking — split long documents for better search precision
• Result type — Pydantic v2 models with Ok/Err pattern
• Input validation — URL, collection names, query length validated

Installation

pip install docs-haven

Or with uv (recommended):

uv pip install docs-haven

Or from source:

git clone https://github.com/Cipher208/docs-haven.git
cd docs-haven
uv sync

With test dependencies

uv sync --extra test

How to Use

1. Add Repositories to Your Knowledge Base

fromstorageimportStoragefrompathlibimportPathstorage=Storage(Path.home() /".docshaven")
# Add a GitHub repo (clones and indexes markdown files)result=storage.add_repo(
url="https://github.com/fastapi/fastapi",
description="FastAPI web framework",
)
print(f"Indexed {result['files_indexed']} files in {result['chunks']} chunks")
# Add with custom file mask (index Python files)result=storage.add_repo(
url="https://github.com/pallets/flask",
mask="**/*.py",
)

2. Search Your Knowledge Base

# Basic searchresults=storage.search("dependency injection")
forrinresults:
print(f"{r['score']:.2f} [{r['collection']}] {r['title']}")
print(f" {r['content'][:100]}...")
print()
# Search with filtersresults=storage.search(
"async middleware",
collections=["fastapi"],
limit=5,
min_score=0.3,
)
# Use auto strategy (fts for short queries, hybrid for long)results=storage.search("how to use Depends()", strategy="auto")

3. Organize with URI Routing

fromuriimportURI, URIRouter# Parse URIsuri=URI.parse("core://fastapi/dependencies")
print(uri.domain) # "core"print(uri.path) # "fastapi/dependencies"print(uri.to_collection()) # "core__fastapi"# Search within a URI scoperouter=URIRouter(storage)
results=router.search_by_uri("core://fastapi", limit=5)
# List all domainsdomains=router.list_all_domains()
# {'core': {'count': 3, 'doc': 'Core documentation'}, 'ref': {'count': 2, ...}}

Available domains:core, ref, guide, lib, src, test, note

4. Detect Conflicts

fromconflictsimportConflictDetectordetector=ConflictDetector(storage)
result=detector.detect(
title="FastAPI dependency injection",
content="How to use Depends()...",
)
ifresult.has_conflicts:
print(f"Found {len(result.candidates)} similar documents:")
forcinresult.candidates:
print(f" - {c['title']} (score: {c['score']})")
print(f" {c['snippet'][:80]}...")
# Record judgmentdetector.judge("new_doc_id", "existing_doc_id", "supersedes")

5. Sync Between Machines

fromsyncimportSyncerfrompathlibimportPathsyncer=Syncer(Path.home() /".docshaven-sync")
# On machine A: exportresult=syncer.export(
{"fastapi": docs, "sqlalchemy": docs},
created_by="alice",
)
print(f"Exported chunk {result['chunk_id']}")
# On machine B: importresult=syncer.import_chunks()
print(f"Imported {result['chunks_imported']} chunks")
# Check statusstatus=syncer.status()
print(f"Chunks: {status['local_chunks']}")

6. Use as MCP Server

Add to your MCP client config (Claude Desktop, Cursor, etc.):

{
"mcpServers": {
"docs-haven": {
"command": "python",
"args": ["/path/to/docs-haven/server.py"]
}
}
}

Then ask your agent:

"Search for FastAPI middleware examples" "What documentation do we have about SQLAlchemy?" "Check if this new doc conflicts with existing ones"

MCP Tools (20)

Tool	Description
kb_search	Search with BM25 ranking + highlighted excerpts
kb_add_repo	Clone and index a GitHub repo
kb_get	Get document content
kb_update	Update document content
kb_delete	Delete document from knowledge base
kb_list_collections	List all collections
kb_stats	Database statistics
kb_collection_rename	Rename a collection
kb_context_add	Add context attachment
kb_context_list	List context attachments
kb_context_rm	Remove context attachment
kb_uri_resolve	URI to collection mapping
kb_uri_search	Search within URI scope (supports wildcards)
kb_uri_list	List URIs in domain
kb_uri_domains	All domains with counts
kb_sync_export	Export compressed chunk
kb_sync_import	Import chunks
kb_sync_status	Sync status
kb_conflict_check	Detect conflicts
kb_conflict_judge	Record judgment

Comparison

Feature	DocsHaven	QMD	Elasticsearch	Context7
Dependencies	1 (mcp)	1 (npm)	JVM + plugins	External service
Setup time	10 seconds	5 minutes	30+ minutes	API key needed
MCP server	Built-in	No	No	Yes
URI routing	Yes	No	No	No
Conflict detection	Yes	No	No	No
Git sync	Compressed chunks	No	No	No
Cost	Free	Free	Free (self-hosted)	Paid tiers

FAQ

What is DocsHaven?

DocsHaven is a local knowledge base designed for AI agents. It provides full-text search via SQLite FTS5, organizes knowledge by URI domains (core://, ref://, guide://), and detects contradictions when adding new documents. It runs as an MCP server with 20 tools.

How is this different from just using SQLite?

DocsHaven adds a complete knowledge management layer on top of SQLite: automatic document chunking, BM25 ranking with LIKE fallback, URI-based organization, conflict detection, and compressed multi-machine sync — all exposed via MCP tools.

Can I use this with Claude Desktop / Cursor / other AI agents?

Yes. DocsHaven runs as an MCP server. Add it to your MCP client config and all 20 tools become available to your agent.

How fast is search?

SQLite FTS5 with BM25 ranking handles 1,000+ documents in under 100ms on modern hardware. No network latency since everything is local.

Is my data sent anywhere?

No. DocsHaven is fully local. The only network operation is cloning GitHub repositories (which you initiate). All search and storage happens on your machine.

Architecture

docs-haven/
├── server.py # MCP server (20 tools)
├── storage.py # SQLite FTS5 backend + Result types
├── vector.py # TF-IDF vector search (optional)
├── uri.py # URI routing
├── sync.py # Git sync (compressed chunks)
├── conflicts.py # Conflict detection
├── result.py # Ok/Err Result type (Pydantic v2)
├── cli.py # CLI interface (11 commands)
├── templates.py # Collection templates
├── import_guard.py # Import validation
├── alias.py # LLM argument aliasing
├── benchmark.py # Performance benchmarks
├── tests/ # pytest test suite (282 tests)
├── docs/ # Documentation + 10 ADRs
└── pyproject.toml # Package config (uv)

Performance

Benchmarked on Linux (Python 3.14, SQLite FTS5):

Operation	Time
Index 1,000 docs	0.076s (13,219 docs/sec)
Search (avg)	3.3ms
Search (P95)	4.3ms
Throughput	299 queries/sec

Run benchmark: uv run python benchmark.py

Integrations

Client	Config	Status
Claude Desktop	claude_desktop_config.json	✅
Cursor	.cursor/mcp.json	✅
Gemini CLI	gemini mcp add	✅
VS Code (Copilot)	.vscode/mcp.json	✅
Codex	.codex/config.toml	✅

Development

# Install with test dependencies
uv sync --extra test# Run tests
uv run pytest tests/ -v
# Run linting
uv run ruff check .# Run formatting
uv run ruff format .# Run type checking
uv run mypy . --ignore-missing-imports
# Run coverage
uv run pytest tests/ --cov --cov-report=term-missing

Security

• SQLite FTS5 with parameterized queries (no SQL injection)
• Input validation on all MCP tool parameters
• No external network dependencies (local-only operation)
• Secret scanning via GitHub Actions (gitleaks)

See SECURITY.md for vulnerability reporting.

Author

Built with ❤️ by Cipher208

License

MIT